On November 4, Tencent released Hunyuan Hy4, the latest version of its flagship LLM series, the Paper and IT Home reported. The company said the Hunyuan (Hy) series has for the first time entered the top tier of open-source models, and that Hy4's weights will be open-sourced soon along with a model system card, with responsible-release measures focusing on cyber and bio-safety risks.
The parameter climb is visible: per Cailian Press' Science Board Daily, the Hy3 series had fewer than 300 billion parameters while Hy4 reaches 770 billion. The Hy4 preview released and open-sourced in August totaled 770B parameters with 49B active and a context window beyond 1M tokens; the formal version significantly expands pre- and post-training scale.

Capability-wise, Hy4 adds visual understanding for the first time and improves instruction following, long-document comprehension, long-horizon agent tasks and deep research. It supports computer use — combining command-line and graphical interfaces — with optimizations for professional software such as Blender and CAD; front-end coding was tuned for complex interactions, business logic and front-back integration so generated pages are both presentable and functional.
Pace matters too. As 21st Century Business Herald observed, Hunyuan is betting this release on office collaboration: the team rebuilt its data and training infrastructure over the past year and now averages a major version every two months.
Hy4 at a glance
Parameters: 770B (Hy3 series: under 300B — more than doubled)
New: visual understanding, computer use (CLI + GUI, Blender/CAD tuning)
Open-source plan: weights soon, system card, cyber & bio-safety focus
Context: 1M (per Hy4 preview)
"Top-tier open-source" is not "top-tier overall." BT Caijing reported that the Hunyuan team openly acknowledges a gap with leading closed models, and its internal bio-safety audit scored Hy4 at just 30.4/100 on open-ended experiment design. The candor matches the responsible-release framing: capability gains and safety assessment now ship on the same schedule.
Open-source competition among Chinese labs has moved from parameter claims to capability and safety claims. Whether Hy4 solidifies its tier will be decided when the weights land and third-party benchmarks re-run.